Future Society Supported by Generative AI
Foundational models accelerating the development of innovative new materials
In ancient times, humans worked with stone, glass, iron, and paper. Since the industrial revolution, they have expanded their search to include ceramics, chemicals, and semiconductors. Mankind has discovered materials with characteristics that are useful in tools and equipment and mastered their use to develop civilization. The development of new materials is an important element in accelerating the advancement of people's lives and society, and many universities, research institutions, and companies are actively and continuously pursuing new materials even today.
As an endeavor that is essential to the future of mankind, the development of new materials is rapidly evolving due to the application of the latest information processing technologies such as generative AI. Materials Informatics ("MI") is becoming widespread as a method for developing new materials using information processing technologies, which is changing the principle of competition in the materials industry to the point where companies are being forced to revamp their business models.
Among the materials developed using MI, there are many that have achieved dramatic improvements in their characteristics. Moreover, MI is expected to help solve social issues such as the protection of the global environment. This is because the development of new materials is essential to realizing solar batteries that convert light into power at high efficiency, storage batteries that are compact, high capacity, and high output, and the generation of photocatalytic hydrogen and ammonia, etc. If MI is effectively utilized, there is a possibility that new materials may quickly be put into practical use to promote the shift to such a decarbonized society.
Advancing the development of new materials while confronting vast possibilities that cannot be fully tested or verified
Until now, the development of new materials was steady, labor-intensive work compared to information technology and other fields of technology development. This is because discovering a new material with the desired properties and establishing a method for creating it required the repetition of time-consuming research, prototyping, and testing. As a result, development periods tended to lengthen, and enormous costs were sometimes required. In extreme cases, a single developer might spend their entire life developing a new material that is never completed, and there are many development projects with hundreds of billions of yen in funding.
One of the reasons why the development of new materials is so challenging is due to the diverse and massive number of parameters that determine the properties of a material. The molecular and crystalline composition and structure are the primary determining factors of the properties of a material, but those properties significantly fluctuate based on environmental factors such as temperature, pressure, the electromagnetic field, etc. In addition, the properties may similarly fluctuate according to the size and shape of the material as well as the presence of defects and impurities in the material. Minor changes in environmental factors and the presence of minute defects can even change the molecular and crystalline structure itself. Such delicate material characteristics are a manifestation of the complex entanglement of phenomena occurring at the atomic scale.
Generally speaking, previous efforts to develop new materials assumed various parameters and a massive number of conditions, which were determined through repeated trial and error to discover a material with the desired properties and establish a way to create it. When doing so, the possibility of obtaining better results increases with more testing of diverse types of parameters and greater precision in the setting range of conditions. However, in reality, definite results must be obtained within a limited time frame and budget. Therefore, it was necessary to narrow down the trial and error conditions to conduct testing and verification.
Developers involved in materials development narrow down the trial and error conditions while referring to results such as document reviews, computer simulations, etc. The documents indicate the conditions that have already been prototyped and verified, and provide many insights that are useful when setting conditions. In addition, running computer simulations can minimize prototyping and verification, with sample preparation requiring vast amounts of time and expense. However, it is up to the developers to decide what types of documents to choose to gain insights and what conditions to set to run a computer simulation. Under such an approach to researching materials, there was a risk of overlooking the potential for superior materials hidden within conditions that were difficult to predict.
In order to develop more useful new materials in a short period of time, materials developers were forced to change how they approached research. MI is a materials development method that holds promise as a new way to streamline development tasks and uncover the potential of useful, new materials hidden within conditions that were previously not tested or verified.
Materials Informatics changing materials development
So how does MI development methodology drive innovation in conventional new materials development?
A massive volume of data has been obtained from the technical literature and experimental results. Moreover, the form of information expression is multimodal (different modes of expression), meaning that it combines text, figures, and images, which makes it difficult to analyze the collected data. MI organizes and standardizes such massive and diverse sets of data and uses advanced information processing technologies such as AI and machine learning, etc. to set targets for the development of effective new materials and explore efficient methods to generate and synthesize those materials.
Applying MI to the development of new materials provides the following advantages.
First, it makes it possible to more efficiently and rapidly narrow down the conditions that should be prototyped and verified based on highly objective insights learned from more information sources than ever before. There are limits to the volume of literature that one developer can read through. Moreover, even if a lot of data was obtained in the past, it is impossible to carefully examine all of it. By using AI and other technologies, it becomes possible to make objective decisions based on statistical processing while referring to literature and data that are too vast for a single developer or development team to handle. This makes it possible to shorten development times and precisely narrow down the conditions for prototyping and verification.
Furthermore, materials developers can explore the potential of unconventional new materials. Information that could lead to the discovery of unexpected new materials may be buried within documents that describe results that received little attention within conventional technology development trends or prototype and experimental data that was abandoned without being deeply explored. With MI, materials developers can examine findings that were not brought to light before and investigate them from an objective perspective. Therefore, MI expands the possibilities for discovering revolutionary new materials.
Using foundational models to rapidly create highly accurate AI models
In addition, there have been efforts in recent years to apply the foundational models that are underlie technology of linguistic generative AI to further streamline the practical implementation of MI. Foundational models are large-scale, general-purpose AI models that provide a foundation for implementing highly accurate AI models for specific applications with minimal training (Figure 1). Generally speaking, the foundational models used in materials development are created by IT companies that can invest huge sums of money and resources into the creation of AI models and collect vast amounts of diverse data about science and technology. User companies customize the highly versatile foundational models that are provided according to their usage objective and utilize them in MI.
For actual analysis and other types of processing, MI uses AI models that are dedicated to specific applications (tasks). For example, when predicting properties from the molecular and crystalline composition and structure, etc., developers prepared and used trained AI models that are dedicated to that task. The reason is that when an AI model that was optimized for a specific task is applied to another task, the analytical accuracy may decrease or in some cases the model may be impossible to apply. When creating each AI model, it was necessary to prepare the appropriate model design and an application-specific data set and train the model. Furthermore, AI training generally requires the execution of a massive amount of computation (deep learning). In other words, each time that an AI model is created for a specific task, a high-performance computer must be prepared to train the model for a reasonably long period of time at great cost. (Upper part of Figure 1)
By contrast, using a foundational model can significantly decrease the data set and the amount of computation during the training required to create an AI model for a specific task. For example, if an AI model for a specific task that was created by training on a data set between several hundred gigabytes and several terabytes using a high-performance GPU cluster (consisting of hundreds to thousands of GPUs) without using a foundational model is instead created based on a foundational model, it can be built with only additional training (fine tuning) on a data set of a few gigabytes using a small number of GPUs.
The impact of applying foundational models to AI model creation in materials development is enormous. As described above, the introduction of MI has made it possible to streamline traditional materials development. However, while the time, effort, and cost required for prototyping and verification have improved, the creation of AI models for specific tasks has in some ways become a factor in producing new inefficiencies. Utilizing foundational models will make it possible to realize the essential goals of MI, which are to shorten the development time for new materials and to enable developers to explore new materials under conditions that are difficult to predict.
By training models on proprietary data collected by companies through prototyping and verification, they can be developed into application-specific forms of AI that acquire material-specific technical know-how. It can be said that continuously fine-tuning application-specific AI models is a technology development asset that will become a source of competitiveness for companies developing new materials. Examples of applying MI using such foundational models to the rapid design of new molecular structures and the development of new materials that combine multiple characteristics are also emerging.
Summary: AI is pioneering next-generation materials development
The environment surrounding materials development is expected to significantly change going forward. Many materials manufacturers are revamping their development systems, collecting experimental data with the intention of using it for AI, and materials developers are increasing their knowledge of data science. Under these circumstances, the practical application of MI using foundational models has the potential to produce spectacular results. It is greatly anticipated that this will accelerate materials development and bring about new changes and new forms of value in our lives and society.